Scrollclaw Score

Virality Scoring

Score every video before publishing. Read references/virality-scoring.md for the full methodology.

Criteria

7 dimensions, scored 0-100 each:

1. Hook Strength — Does frame 1 + first 3 seconds force a pause?
2. Emotional Impact — Does it trigger a visceral response (not just "interesting")?
3. Pacing — Does every second earn the next second?
4. Text Readability — Are captions legible on mobile, properly timed?
5. Scroll-Stop Power — Would this stop a thumb mid-scroll in a feed?
6. Completion Likelihood — Will viewers watch to the end?
7. Shareability — Would someone send this to a friend?

Thresholds

Process

1. Watch the final video on a phone screen, in-app, scrolling past it like a user
2. Score each dimension independently
3. Calculate average
4. If below threshold, identify the weakest dimensions
5. Route back to the appropriate skill:
   • Hook/scroll-stop weak → /first-frame (regenerate canonical face)
   • Pacing/completion weak → /persona (restructure script)
   • Emotional impact weak → /persona (better language from research)
   • Text readability weak → /assemble (fix captions)
   • Audio/voice issues → /assemble (re-run S2S)

Brand Memory Integration

Reads

Writes

Context loading

Copy code
📊 Score context loaded:
  ✓ Campaign: ridge-q1
  ✓ Final clips: 2 (workspace/campaigns/ridge-q1/clips/final-*.mp4)
  ✓ Brief: workspace/campaigns/ridge-q1/brief.md
  ✓ Persona research: workspace/campaigns/ridge-q1/persona-research.md

Learnings append format

After every score, append to workspace/campaigns/<slug>/learnings.md:

markdownCopy code
## YYYY-MM-DD — <campaign-slug> <format>

**Score:** <average>/100 (<go/no-go>)
**Dimension breakdown:** Hook <n> | Emotional <n> | Pacing <n> | Captions <n> | Scroll-stop <n> | Completion <n> | Share <n>

**What worked:**
- [specific finding]

**What didn't:**
- [specific finding]

**Remediation:** [if no-go: which skill to route back to and why]

Never overwrite this file. Only append.

Contract

Input

• Required: at least one final assembled clip
• Optional: campaign brief, persona research, prior campaign learnings
• Format: workspace video files plus markdown context
• Source: /assemble, workspace/campaigns/<slug>/brief.md, and persona-research.md

Output

• Produces: one score card per clip, a go/no-go decision, and appended learnings
• Format: markdown files in workspace/campaigns/<slug>/scores/ plus append-only entries in learnings.md
• Default behavior: score every final clip against the 7 dimensions and route the user back to the failing stage if the score is below 70
• Downstream use: publish/no-publish decision and future campaign learnings

Validation

• Pre-conditions: final clip exists and is watchable end to end
• Post-conditions: score includes a dimension breakdown, a threshold decision, and clear remediation if needed
• Failure checks: do not recommend publishing without a real breakdown; if the clip is unscorable or missing context, flag the gap explicitly

Output

• Virality score card in workspace/campaigns/<slug>/scores/score-<version>.md
• Go/no-go decision
• Learnings appended to workspace/campaigns/<slug>/learnings.md
• If no-go: specific remediation routing

Batch Scoring

For campaign batches, score all variants and rank. Publish only the 70+ tier. Below-60 variants get regenerated, not tweaked.

bashCopy code
bash scripts/batch-campaign.sh \
  --workspace workspace/campaigns/<slug> \
  --creators workspace/campaigns/<slug>/creators/ \
  --formats "talking-head,pov-demo,podcast-clip" \
  --dual-output